Document Type

Conference Proceeding

Publication Date


Publication Title

Medical Physics


Purpose: Currently there is no validated method for quantifying risk of disease progression in chronic obstructive pulmonary disease (COPD). We aim to address this by predicting whether a patient will worsen using a machine learning model trained on basic patient information, current pulmonary function values, FEV1 and FVC, and extracted features from a non-contrast CT scan. Disease severity is characterized by these pulmonary function values, and by extension disease progression can be determined by the change in these values between future timepoints. Methods: XGBoost, a popular classification library that utilizes gradient boosted trees, was used to define an ensemble model composed of multiple decision trees with limited depth whose individual predictions are summed to yield final predictions. A binary classification model was used to predict the probability of a patient’s FEV worsening by 200 mL in five years. The model was trained on data from 400 COPDgene patients, including age, BMI, gender, current FEV1 and FVC value, and mean lung density within each lung lobe at inhale and exhale (as measured on non-contrast inhale/exhale CT). Feature importance was calculated using Shapley values. Results: A 10-fold cross-validation train-validation-test split yielded a mean ROC-AUC of .73 and .63 for validation and test sets, respectively, for predicting five-year future FEV1 value. Initial FEV1 and FVC were calculated to be the most important features with Shapley Values of .016 and .012, contributing to 24.5% and 18.5% of model output, respectively. Conclusion: Results indicate that CT density measurements and patient information can be used to identify patients at risk of FEV decline. Predictive capability can be further improved with more sophisticated deep learning approaches and including more patient data.





First Page


Last Page



American Association of Physicists in Medicine 65th Annual Meeting & Exhibition, July 23-27, 2023, Houston, TX